Employing Event Inference to Improve Semi-Supervised Chinese Event Extraction

Proceedings of COLING 2014, the 25th International Conference on Computational Linguistics: Technical Papers, pages 2161–2171, Dublin, Ireland, August 23-29 2014. Employing Event Inference to Improve Semi-Supervised Chinese Event Extraction Peifeng Li, Qiaoming Zhu, Guodong Zhou School of Computer Science & Technology Soochow University, Suzhou, 215006, China {pfli, qmzhu, gdzhou}@suda.edu.cn Abstract Although semi-supervised model can extract the event mentions matching frequent event patterns, it suf- fers much from those event mentions, which match infrequent patterns or have no matching pattern. To solve this issue, this paper introduces various kinds of linguistic knowledge-driven event inference mechanisms to semi-supervised Chinese event extraction. These event inference mechanisms can capture linguistic knowledge from four aspects, i.e. semantics of argument role, compositional semantics of trig- ger, consistency on coreference events and relevant events, to further recover missing event mentions from unlabeled texts. Evaluation on the ACE 2005 Chinese corpus shows that our event inference mech- anisms significantly outperform the refined state-of-the-art semi-supervised Chinese event extraction system in F1-score by 8.5%. 1 Introduction An event is a specific occurrence involving arguments (participants and attributes) of the specific roles. In an event, trigger is the main word which most clearly expresses its occurrence, so recognizing an event can be recast as identifying a corresponding trigger. An event may have several arguments, which are entity mentions (e.g., person name, time, location, etc.) and must fulfill the corresponding roles. Take the following sentence as an example: S1: On the 25th Dec. (A1: Artifact), peacekeepers (A2: Artifact) returned (E1: Transport) to Am- man (A3: Place) by flight (A4: Vehicle). For this example, an event extraction system should identify one event mention E1, which is trig- gered by verb “returned” whose event type is Transport, with four arguments, “peacekeepers”, “25th Dec.”, “flight”, and “Amman”, fulfilling the roles of Artifact, Time, Vehicle, and Place, respectively. Automatically extracting events from free texts is a higher-level Information Extraction (IE) task, which is still a challenge due to the complexity of natural language and the domain-specific nature, especially in Chinese for its specific characteristics. In particular, most of previous studies have fo- cused on English event extraction, while only a few concern Chinese. Currently, supervised learning models have dominated event extraction. To reduce the labeled data required, a few semi-supervised models have been applied to English event extraction (e.g., Riloff 1996; Yangarber et al., 2000; Stevenson and Greenwood, 2005; Huang and Riloff, 2012). Since classi- fier-based model needs dozens of annotated documents to train model, most of previous semi- supervised models focused on pattern-based approach, which only needed a few seed (event) patterns. In those pattern-based approaches, frequent event patterns, which occur in many documents, were chosen as relevant patterns to match event mentions in unlabeled texts. However, the order of words in a Chinese sentence is rather agile for its open and flexible structure, and different orders might express the same meaning due to the semantics-driven nature of the Chinese language. This results in the di- versity of Chinese event patterns and numerous infrequent patterns, …

Content truncated to 500 words